Microsoft Unveils Maia 300 Chip: 2nm Technology and Reduced Dependence on Nvidia

On October 12, 2026, Microsoft unveiled the third generation of its custom artificial-intelligence chip family — Maia 300. According to Reuters, the company plans to deploy the chip broadly across its cloud data centers in 2027.

Under the announced plan, Maia will become the primary AI inference system for Microsoft's products and services. That is, running the corporation's artificial-intelligence models — the process of responding to user queries — will be handled mainly by these custom chips.

The Maia family is a custom silicon platform designed for Microsoft's internal AI services. The unveiling of the third generation shows the company continuing to invest in its own chip architecture. The previous generation — Maia 200 — is now being replaced by the new chip.

The Maia family is tailored to the needs of Microsoft's cloud platform: the chips power artificial-intelligence services running in the company's own data centers. The arrival of the third generation shows that work in this area is not a one-off experiment but a long-term program.

Key aspects of the announcement

Maia 300 is Microsoft's third-generation custom AI chip. It was unveiled as the successor to the company's previous generation — Maia 200. The presentation took place on October 12, 2026, and immediately drew industry attention: it is not just about a new chip, but about a plan to deploy more than a million units.

According to Reuters, Microsoft will deploy the chip at scale across its own cloud infrastructure in 2027. This is not a mass-market release: the company will use it internally within its own data centers. This model is common among major cloud providers — custom chips are first tailored to the needs of their own infrastructure and are not released to the external market.

The Verge focused on the technical aspects of the announcement: the new generation is significantly ahead of its predecessor in energy efficiency and throughput. Both sources agree on one thing — this is one of the biggest steps in Microsoft's own silicon strategy.

Inference: the chip's primary job

Maia 300 is designed specifically for inference workloads. Artificial-intelligence systems have two main computing stages: training and inference.

Training is the process of teaching a model: it processes large volumes of data and learns the patterns within it. This is a one-time but extremely compute-heavy stage — training the largest models takes weeks across thousands of chips.

Inference, meanwhile, is the trained model's daily work: when a user sends a query, the model returns an answer. Every conversation, every text or image query is a separate inference operation. A model is trained once but then answers millions of users' queries every day, every hour.

For cloud providers, inference is what demands the largest total compute volume. That is why the efficiency of inference chips directly affects service cost and capacity. Maia 300's announced metrics — more useful compute per megawatt and higher token throughput — are aimed squarely at this task: processing large query volumes with minimal energy consumption.

There are two ways to improve inference efficiency: reduce the time per query and increase the number of queries processed simultaneously. Maia 300's announced metrics emphasize the second direction: a sevenfold increase in rack-scale token throughput raises the number of users served at the same time.

Efficiency: 40 percent and sevenfold gains

Maia 300's announced technical specs focus on two main areas: energy efficiency and throughput.

The company notes that Maia 300 delivers approximately 40 percent more useful compute per megawatt than leading GPU-based systems. This metric matters for data centers: the same electricity consumption can power more artificial-intelligence work. Measuring per megawatt is directly tied to data centers' main constraints — electrical power and cooling capacity: when a facility's power is limited, the return on every watt becomes decisive.

It also matters that the 40 percent figure is framed as “useful compute”: this is not about theoretical peak power but about practical performance under real inference workloads. This measurement method reflects the chip's efficiency not in lab conditions but in actual data-center operation.

The second metric concerns rack-scale throughput. A full rack of 96 Maia 300 accelerators delivers up to seven times higher rack-scale token throughput than the previous Maia 200 generation. Token throughput is the volume of data an artificial-intelligence model can process per unit of time: the higher this figure, the more user queries the system can serve simultaneously.

According to Microsoft, a full rack of 96 Maia 300 accelerators delivers up to seven times higher rack-scale token throughput than the previous Maia 200 generation (The Verge).

A sevenfold increase is a major leap within a single generation, and it comes not only from the speed of an individual chip but from the architecture at the level of the entire rack — the interconnects between chips and the efficiency of their joint operation. Rack-scale measurement matters precisely because modern AI workloads run not on a single chip but in the coordinated work of dozens and hundreds of chips.

2nm chiplets and rack architecture

Maia 300 is built on a chiplet architecture. This means the chip consists not of one giant die but of a set of interconnected smaller dies (chiplets). This approach reduces the manufacturing challenges typical of large chips: smaller dies are easier to produce at high quality, and they are then combined into a single system through high-speed interconnects.

The chiplets designed for the main compute tasks are manufactured on TSMC's 2-nanometer process. Two nanometers is one of the most advanced volume chip-manufacturing process technologies currently available. A smaller process node allows more transistors to be placed in the same area, delivering higher performance and lower power consumption. TSMC is the leading contract manufacturer of advanced chips, and the move to a 2-nanometer process is considered one of the industry's latest technological milestones.

Another advantage of the chiplet approach is scalability: chiplets for different tasks can be combined on a single platform. For example, while the main compute chiplets are produced on the most advanced — 2-nanometer — process, auxiliary blocks can be made on cheaper, more mature technology. This optimizes both costs and manufacturing efficiency.

At the rack level, the architecture is as follows: a single rack holds 96 Maia 300 chips, with total power consumption of about 167 kilowatts. This is a high-density computing system: a complete artificial-intelligence inference cluster fits inside one rack. 167 kilowatts is a lot of power for a single rack — such density requires dedicated power and cooling infrastructure, but in return an enormous amount of compute power is concentrated in one rack. For comparison: an ordinary server rack typically draws around 10–20 kilowatts, meaning a Maia 300 rack consumes nearly ten times as much power. That is why systems of this density require dedicated power and cooling designs in data centers.

One million chips: the scale of deployment

One of the most striking aspects of the announcement is the scale of deployment. Microsoft plans next year to install more than a million Maia 300 chips for its internal artificial-intelligence services.

This figure shows the chip is becoming not an experimental project but a core element of the company's infrastructure. Large-scale deployment will begin in 2027 and will cover the company's cloud facilities.

One million chips means computing power equivalent to tens of thousands of racks: with 96 chips per rack, a million chips would require more than ten thousand full racks. Such scale also shows how vast the company's data-center network is — the chips will be distributed not across a single facility but throughout the cloud infrastructure.

Designating Maia as the primary AI inference system will affect Microsoft's broad product range: artificial-intelligence models will run on these chips across all lines — from cloud services to consumer products. This is the next stage in the history of the company's own chip strategy: following the previous generation — Maia 200 — the third generation arrives alongside large-scale deployment, showing the company steadily continuing its silicon path.

Reduced dependence on Nvidia

As Reuters notes, this move reduces Microsoft's long-term dependence on Nvidia. Nvidia currently holds the leading position in the artificial-intelligence chip market, and many cloud providers rely on its GPUs.

Developing its own custom chip and deploying it in the millions gives Microsoft the ability to manage its computing infrastructure independently. The company is taking a step toward securing the capacity needed for internal AI services without dependence on external suppliers.

Economically, this is vertical integration: the chain of designing the chip, ordering its production, and running it in the company's own data centers will be under the control of a single company. This model reduces dependence on an external supplier's pricing policies and supply constraints, and in the long term allows the cost of computing to be lowered. According to Reuters, Microsoft's large-scale deployment plan for its own chips also serves a cost-saving goal: a company with its own silicon can directly control the cost of computing capacity.

The Maia 300 unveiling was the next stage in the trend of large technology companies moving to custom silicon. Every major cloud provider is now developing its own chip architecture, and Microsoft's third-generation chip is one of the latest major announcements in this area.